Application of probabilistic modelling analysis and filtering into image segmentation

In the generation of volume rendered images, Transfer Function (TF) specification has an important role. The specification of an appropriate TF that allows intuitive control of the visual parameters is a very challenging task. In Magnetic Resonance Imaging (MRI), this process is even harder since certain types of volume data are not standardized in pre-defined range of data values. Based on the MRI sequence in use and depending on the acquisition parameters, data specific sub-ranges might be assigned to the same type of structure. To be able to work in all possible cases, automatic sub-range detection methods are required. In this study, a semi-automatic method for initial generation of TFs is introduced. Our approach is based on Volume Histogram Stack (VHS) which is created by re-generating the image slices of a dataset based on a user specified spline. VHS is obtained by aligning the histograms of the image slices of the MRI series. By considering the VHS structure, Gaussian basis functions provide a good approximation for fitting the lobes of the VHS. Approximating the VHS using Gaussian basis functions allows a coarse classification and enables an effective initial TF design. The developed technique employs hierarchical approximation of the VHS using Gaussians with multiple orientations and scales. Then, a finer classification step is carried out for refinement of the initial result. As a finer classification, which is based on spatial domain knowledge, such morphological operations as dilation, and erosion and region growing are applied. The proposed method is applied to 29 (14 T1 DUAL+10 T1 WATS+5 THRIVE) MRI datasets for abdominal tissue/organ visualization. The results show that the proposed system provides a useful and intuitive initialization for TF design. Applications to several MRI datasets testify the success of the developed technique in accurate visualization of abdominal tissues/organs.